SARM2 arrives as a stage-aware reward model that feeds dense signals into the SPIRAL loop, letting robots refine their own vision-language-action policies through autonomous rollouts instead of constant human labeling.
SARM2 is a multi-task, stage-aware reward model: dense, accurate AND general.
Feeding SARM2 rewards into a self-improvement framework SPIRAL boosts task success rates from 50-60% to 90-100% 🔥
Congrats to @QianzhongChen (XDOF intern) and co-authors: - @BrianZheng103 - @uynitsuj (XDOF intern) - @suning_huang - @JiankaiSun
Folding shorts climbs from 58 percent success to a clean 100 percent while whiteboard cleaning rises from 50 percent to 90 percent, with overall task rates crossing 90 percent on the ten-task benchmark.
Code, weights, and any hardware integration plans have not been confirmed yet, so the timeline for external labs or production robots remains open.
Users are excited about the SARM2 reward model's self-improving loops for robot task learning because they see the concept and execution as outstanding and massive for advancing capabilities.
No Digg Deeper questions have been answered for this story yet.
Training action-primitive-specific reward models is a great idea! + nice system acronym: 🌀Thank you Qianzhong and team @xdofai
Introduce SARM2 🤖a multi-task stage-aware reward model that empowers a self-improving loop: 🧺 Folding Shorts 58% → 100% 🧽 Cleaning Whiteboard 50% → 90% Paper + project page below 👇 (1/n)
Read more about SARM2 on the XDOF blog: https://xdof.ai/blog/sarm2
Check out the SARM2 project page: https://qianzhong-chen.github.io/sarm2.github.io/
@xdofai @QianzhongChen @BrianZheng103 Heyyyy, follow me bck please, let's pump it
@xdofai Looking forward to hearing more from the team.
@xdofai @QianzhongChen @BrianZheng103 Self-improvement loops like this are massive for helping robots master
@xdofai @QianzhongChen @BrianZheng103 Excellent project, happy to connect.
@Ken_Goldberg @xdofai Than you Ken!
@xdofai Outstanding concept and execution, happy to collaborate.
@xdofai Got a growth idea for your project. Open to chat?
